As high computing power is available at an affordable cost, we rely on microprocessor-based systems for much greater variety of applications. This dependence indicates that a processor failure could have more diverse impacts on our daily lives. Therefore, dependability is becoming an increasingly important quality measure of microprocessors.;Temporary hardware malfunctions caused by unstable environmental conditions can lead the processor to an incorrect state. This is referred to as a transient error or soft error. Studies have shown that soft errors are the major source of system failures. This dissertation characterizes the soft error behavior on microprocessors and presents new microarchitectural approaches that can realize high dependability with low overhead.;Our fault injection studies using RISC processors have demonstrated that different functional blocks of the processor have distinct susceptibilities to soft errors. The error susceptibility information must be reflected in devising fault tolerance schemes for cost-sensitive applications. Considering the common use of on-chip caches in modern processors, we investigated area-efficient protection schemes for memory arrays. The idea of caching redundant information was exploited to optimize resource utilization for increased dependability. We also developed a mechanism to verify the integrity of data transfer from lower level memories to the primary caches. The results of this study show that by exploiting bus idle cycles and the information redundancy, an almost complete check for the initial memory data transfer is possible without incurring a performance penalty.;For protecting the processor\u27s control logic, which usually remains unprotected, we propose a low-cost reliability enhancement strategy. We classified control logic signals into static and dynamic control depending on their changeability, and applied various techniques including commit-time checking, signature caching, component-level duplication, and control flow monitoring. Our schemes can achieve more than 99% coverage with a very small hardware addition. Finally, a virtual duplex architecture for superscalar processors is presented. In this system-level approach, the processor pipeline is backed up by a partially replicated pipeline. The replication-based checker minimizes the design and verification overheads. For a large-scale superscalar processor, the proposed architecture can bring 61.4% reduction in die area while sustaining the maximum performance.
展开▼
